[published in Interfaces, 31:3, Part 2 of 2, May June 2001, pp. S56-S73.

Thirty Years of Conjoint Analysis: Reflections and Prospects

Paul E. Green Suite 1400, Steinberg Hall-Dietrich Hall The Wharton School University of Pennsylvania Philadelphia, PA 19104-6371

Abba M. Krieger Suite 3000, Steinberg Hall-Dietrich Hall The Wharton School University of Pennsylvania Philadelphia, PA 19104

Yoram [Jerry] Wind Suite 1400, Steinberg Hall-Dietrich Hall The Wharton School University of Pennsylvania Philadelphia, PA 19104-6371

2

About the Authors Paul Green has been in or around marketing research for over fifty years, both in industry and academia. He’s still tolerated by his younger colleagues. His off-hours are spent either on the ski slopes or playing the piano at local nursing homes [he’s too old to play in bars anymore]. Abba Krieger is equally at home in statistics, operations research, computer science, and marketing. He is clearly a man for all seasons. He’s a devoted family man with five children, who somehow manages to spend quality time at home despite a hectic teaching schedule and extensive service to the Wharton School. Jerry Wind is also multi-talented – academic, consultant, educator, and all-round catalyst for developing new academic programs and initiating new directions for the Wharton School. No one can figure out how he seems to thrive on 18-hour workdays. Somehow he still manages to play tennis regularly and see the latest Broadway shows and art exhibitions.

3

Abstract Conjoint analysis is marketers’ favorite methodology for finding out how buyers make trade-offs among competing products and suppliers. Conjoint analysts develop and present descriptions of alternative products or services that are prepared from fractional factorial, experimental designs. They use various models to infer buyers’ partworths for attribute levels. and enter the partworths into buyer choice simulators to predict how buyers will choose among products and services. Easy-to-use software has been important for applying these models. Thousands of applications of conjoint analysis have been carried out over the past three decades.

Keywords: choice models; measurement; new products, scaling methods

4

Thirty Years of Conjoint Analysis: Reflections and Prospects

Introduction

Readers of Interfaces are no strangers to multiattribute utility models [Keeney and Raiffa 1976; Starr and Zeleny 1977; Bell, Raiffa, and Tversky 1988]. Thomas Saaty [1980] introduced a different approach to multiattribute utility measurement: the analytic hierarchy process [AHP]. Both approaches emphasized small numbers of decision makers facing high-level decisions. Operations researchers and management scientists have applied both methods extensively to important problems in management and government decision making. OR/MS researchers may be less familiar with another method – conjoint analysis – that has been applied for over 30 years, primarily by researchers in marketing and business. Conjoint analysis evolved from the seminal research of Luce and Tukey [1964]. Their theoretical contributions were put to use by a number of psychometricians, including Carroll [1969], Kruskal [1965], and Young [1969]. These researchers developed a variety of nonmetric models for computing partworths (attribute-level values) from respondents’ preference orderings across multiattributed stimuli, such as descriptions of products or services. Conjoint analysis is, by far, the most used marketing research method for analyzing consumer trade-offs. Surveys conducted by Wittink and Cattin [1989] and Wittink, Vriens, and Burhenne [1994] attest to its world-wide popularity. It is not difficult to see why researchers developed and applied conjoint analysis so rapidly. Conjoint analysis deals with a central management question: Why consumers choose one brand or one supplier over another? Also, marketing research practitioners want to be part of something new, and computer software for implementing the methodology became readily available.

The Basic Ideas of Conjoint Analysis

Conjoint analysis is one of many techniques for handling situations in which a decision maker has to deal with options that simultaneously vary across two or more attributes. The problem the decision maker faces is how to trade off the possibility that option X is better than option Y on attribute A while Y is better than X on attribute B, and various extensions of these conflicts.

Conjoint analysis concerns the day-to-day decisions of consumers – what brand of

toothpaste, automobile, or photocopying machine to buy [or lease]? Marketing researchers may collect trade-off information for hundreds or even thousands of respondents. Data collection and processing techniques must be fairly simple and routinized to handle problems of this scope.

Conjoint analysis is a technique for measuring trade-offs for analyzing survey responses

concerning preferences and intentions to buy, and it is a method for simulating how consumers might react to changes in current products or to new products introduced into an existing

5

competitive array. Researchers have used conjoint analysis for consumer and industrial products and services and for not-for-profit offerings.

To understand the basic concepts of conjoint analysis, assume that a marketer of credit

cards wishes to examine the possibility of modifying its current line of services. One of the first steps in designing a conjoint study is to develop a set of attributes and corresponding attribute levels to characterize the competitive domain. Focus groups, in-depth consumer interviews, and internal corporate expertise are some of the sources researchers use to structure the sets of attributes and levels that guide the rest of the study.

In an actual study of credit-card suppliers, researchers used a set of 12 attributes with two

to six levels, for a total of 35 levels (Table 1). However, the total number of possible combinations of levels is 186,624. Conjoint analysts make extensive use of orthogonal arrays [Addelman 1962] and other types of fractional factorial designs to reduce the number of stimulus descriptions that a respondent sees to a small fraction of the total number of combinations. In this problem, an array of 64 profiles [less than 0.04 percent of the total] is sufficient to estimate all attribute-level main effects on an uncorrelated basis. Since the study designers used a hybrid conjoint design [Green and Krieger 1996], each respondent was asked to consider only eight (balanced) profile descriptions drawn from the 64 profiles.

Table 1: These attributes (and levels within attributes) describe the set of potential services that would be offered to credit card subscribers.

Annual price [$] 0 10 20 50 80 100 Cash rebate [end-of-year, on total purchases] None ½% 1% 800 number for message forwarding None 9-5 weekdays 24 hours per day Retail purchase insurance None 90 days’ coverage Common carrier insurance [death, injury] None $50,000 $200,000 Rental car insurance [fire, theft, collision, vandalism] None $30,000 Baggage insurance None $2,500 depreciated cost $2,500 replacement cost Airport club admission [based on small entrance fee] No admission $5 per visit $2 per visit Card acceptance

Air, hotel, rental cars [AHC] AHC and most restaurants [AHCR] AHCR and most general retailers [AHCRG] AHCR and department stores only [AHCRD] 24-hour medical/legal referral network No Yes Airport limousine to city destination Not offered Available at 20% discount 800 number for emergency car service Not offered Available at 20% discount

6

For such studies, researchers may prepare prop cards (Figure 1). After the respondent sorts the prop cards in terms of preference, each card is rated on a 0 to 100 likelihood-of-acquisition scale. In small conjoint studies (for example, six or seven attributes, each at two or three levels), respondents are given all of the full profiles – 16 to 32 prop cards. In these cases, respondents typically sort the prop cards into four to eight ordered categories before they give likelihood-of-purchase ratings for each separate profile within each category.

Figure 1: These prop cards illustrate specific services that a credit card could offer. For each card the respondent indicates how likely she would be to subscribe to the credit card on a 0-100 point scale.

1Annual price$20

Cash rebateNone

800 number for message forwardingNone

Retail purchase insuranceNone

Common carrier insurance$50,000

Rental car insurance$30,000

Baggage insurance$2,500 depreciated cost

Airport club admission$2 per visit

Card acceptanceAir, hotel, rental cars

Medical-LegalNo

Airport limousineNot offered

800 number for emergency carAvailable at 20% discount

2Annual price$50

Cash rebate½ %

800 number for message forwarding24 hours per day

Retail purchase insuranceNone

Common carrier insuranceNone

Rental car insurance$30,000

Baggage insuranceNone

Airport club admission$5 per visit

Card acceptanceAir, hotel, rental cars and most restaurants

Medical-LegalYes

Airport limousineAvailable at 20% discount

800 number for emergency carNot offered

Types of Conjoint Data Collection Four major types of data collection procedures are currently used for conjoint analysis:

(1) In full profile techniques, each respondent sees a complete set of the full-profile prop cards. After sorting the cards into ordered categories, the respondent rates each card on a 0 to 100 likelihood-of-purchase scale.

(2) In compositional techniques, such as the CASEMAP procedure [Srinivasan

1988], each respondent rates the desirability of each set of attribute levels on a 0 to 100 scale and then rates the attributes on an importance scale. (This approach is typically called self-explicated preference-data collection.)

7

(3) In hybrid techniques, each respondent performs a self-explicated evaluation task and evaluates a subset of the full-profile cards [Green, Goldberg and Montemayor 1981]. The resulting utility function is a composite of data obtained from both tasks.

(4) In adaptive conjoint analysis, a hybrid technique developed by Sawtooth Software

[Johnson 1987], each respondent first performs a self-explication task and then evaluates a set of partial-profile descriptions, two at a time. Partial profiles usually consist of two or three attributes per stimulus card. Researchers vary the partial profile descriptions depending upon responses to earlier paired comparisons. The respondent evaluates each pair of partial profiles on a graded, paired comparisons scale. Both tasks are administered by computer [Johnson 1987].

Conjoint Models Most conjoint analysts fit what is known as the part-worth model to respondents’ evaluative judgments, whether they obtain these judgments using full-profile, self-explicated, or hybrid approaches. However, they occasionally use vector and ideal point models. We assume that there are P attributes and J stimuli used in the study design. For a given respondent, we let yjp denote the desirability of the pth attribute for the jth stimulus; we first assume that yjp is inherently continuous. The vector model assumes that the respondent’s preference sj for the jth stimulus is given by

∑=

=P

pjppj yws

1

where wp denotes the respondent’s importance weight for each of the P attributes (Figure 2). Figure 2: These diagrams illustrate what is meant by linear preferences, ideal point preferences, and discrete (part-worth) preferences. The third graph shows three partworths.

Preference

Amount of attribute p

Increasing preference

Decreasing preference

Amount of attribute p Chosen levels of attribute p

Ideal point xp

Highest preference

Lowest

Highest preference

Nexthighest

preference

Source: Green, P.E., and Srinivasan, V. [1978], Journal of Consumer Research, 5, 103-123.

8

In the ideal-point model, the analyst posits that preference sj is inversely related to the weighted squared distance 2

jd of the location yjp of the jth stimulus from the individual’s ideal

point xp, where 2jd is defined as

( )∑=

−=P

ppjppj xywd

1

22 .

In the part-worth model, the analyst assumes that

( )∑=

=P

pjppj yfs

1

where yjp is the category level and fp is a function denoting the part-worth corresponding to level yjp. In practice, the analyst estimates fp(yjp) for a selected set of discrete levels of yjp. In Figure 3, we show illustrative (averaged) part-worths for the attribute levels described in Table 1. Part-worths are often scaled so that the lowest part-worth is zero, within each attribute. Strictly speaking, analysts evaluate part-worth functions at discrete levels for each attribute. However, in many applications, analysts interpolate between levels of continuous attributes, such as price, when they enter the part-worths in buyer-choice simulators. The scaling (vertical axis) is common across all attributes; this allows the analyst to add up part-worths across each attribute level to obtain the overall (product or service) utility of any profile composable from the basic attribute levels. Figure 3: This chart illustrates how part-worth scale values depend on price, message forwarding, … emergency car service. For example, the most preferred price is $0 and the least preferred price is $100.

0.5

0.4

0.3

0.2

0.1

ScaleValues

Price

$0$10

$20

$50

$80 $100 None½%

1%

Rebate

None9-5 pm

24 hours

MessageForwarding

None

90 days

None

$50k

$200k

None

$30k

PurchaseInsurance

CarrierInsurance

CarInsurance

BaggageInsurance

None

$2.5k(D)

$2.5k(R)

None$5

AirportClubs

AHCAHCR

AHCRG

Acceptance

No

Yes

No

Yes

Medical/Legal

Limousine EmergencyCar Service

$2

AHCRDNo

Yes

Stimulus Presentation In collecting conjoint data analysts currently emphasize the full-profile and hybrid procedures, including Sawtooth’s adaptive conjoint analysis. While they still employ paragraph descriptions of attribute levels in some industry studies, they usually use profile cards with terse

9

attribute-level descriptions (Figure 1). Analysts increasingly use pictorial material; these kinds of props make the respondent’s task more interesting and convey information easily with little ambiguity. Moreover, conjoint methods are increasingly being applied to the design of physical products (for example, foods and beverages, fragrances, and personal-care products). In these cases, researchers use actual, experimentally designed prototypes.

Precursors to Conjoint Analysis

How did conjoint analysis come about? For marketing researchers working in data analysis and modeling, the 1970s provided a rich bounty of tools and techniques. Unlike the 1960s (a decade of borrowing tools from operations researchers), the late ‘60s and early ‘70s saw the strong influence of developments in the behavioral sciences (primarily psychometrics and mathematical psychology) on marketing research methods. In particular, three techniques, cluster analysis, multidimensional scaling [MDS], often called perceptual and preference mapping, and conjoint analysis were introduced to marketing at that time. Cluster analytic methods almost immediately found application in market segmentation [Green, Frank, and Robinson 1967]. Unlike a priori market segmentation, where researchers assume they know segment identifiers at the outset, cluster analysis provided researchers and practitioners with tools for implementing post hoc, or cluster-based segmentation. Researchers could then base segmentation on the needs or benefits sought, brand preferences, problem-solving alternatives, psychographics, and a host of other variables. The idea of cluster-based segmentation was to let the data speak for themselves in terms of finding groups of consumers who share similar needs, attitudes, trade-offs, or benefits. Multidimensional scaling methods also received considerable attention in the 1970s. Marketing researchers learned the value of constructing attractive “maps” for pictorially representing large two-way (or multiway) numerical tables as point or point and vector geometric representations in two (or possibly three) dimensions. They could also profitably combine MDS with hierarchical clustering methods to augment two-dimensional configurations with cluster-based representations based on solutions in higher dimensionalities. Carroll and Green [1997] summarize the history of the MDS field. As early appliers of MDS found (often to their dismay), the brand-positioning maps generated using MDS techniques appeared to be largely of diagnostic rather than predictive value. Researchers who wanted to use MDS to develop new or restaged products faced a two-part problem: translating brand scores on perceived dimensions to manipulable dimensions, and relating manipulable attribute levels to their counterparts in perceptual or preference space.

These problems illustrate the general problem of reverse engineering, in which back-translation from perceived attributes to physical and chemical product characteristics is typically not one-to-one [Kaul and Rao 1994]. It is little wonder that conjoint analysis, with its emphasis on researcher-specified attribute levels, provided a basis for relating preference or choice responses to explicit arguments, including physical attributes as well as verbal descriptions of product and service attribute levels.

10

The Era of Conjoint Analysis

Conjoint analysis has been blessed with several sets of parents. The seminal precursory paper to conjoint analysis, Luce and Tukey’s paper on conjoint measurement, appeared in the Journal of Mathematical Psychology in 1964. The authors focused on axiomatic approaches to fundamental measurement. The idea was to obtain ordered metric scale data from rank-order response data and a set of factorially designed stimuli. Not surprisingly, the initial conjoint algorithm, called MONANOVA, designed by Kruskal [1965] and programmed by Kruskal and Carmone, used ranked response data. In the late 1960s, Green and his colleagues (Vithala Rao, Frank Carmone, and Arun Jain) started running numerous experiments with the MONANOVA program. A working paper by Green and Rao appeared in 1969, followed by the first marketing journal article on conjoint analysis [Green and Rao 1971]. Following this, Johnson [1974] and Westwood, Lunn, and Beazley [1974] published articles on Johnson’s two-attributes-at-a-time trade-off model. Users of Kruskal’s MONANOVA and Shocker and Srinivasan’s [1977 ] LINMAP programs [both nonmetric models] soon learned, however, that ratings-based dependent variables analyzed with dummy-variable regression techniques provided a robust alternative to ordinally based data procedures [Cattin and Wittink 1976; Carmone, Green and Jain 1978]. Orthogonal main-effects plans, based on Addelman’s [1962] fractional factorial designs, significantly reduced respondents’ cognitive burdens in responding to full-profile descriptions. Hence, in the mid-‘70s conditions were ripe for the quick diffusion of metric methods of conjoint analysis, using dummy-variable regression. In the mid-‘80s, Johnson [1987] introduced his adaptive conjoint analysis program that used graded paired comparisons as one set of inputs to the model. About the same time, Herman [1988] introduced a PC-based package that used full profile stimuli based on orthogonal designs. Both packages contained conjoint simulators. Both software packages were also easy to use and moderately priced for commercial research firms. The advent of PC-based conjoint packages opened a large and eager market for applying the methodology. By the late ‘70s, conjoint analysis had truly come of age. Wittink and his co-authors [Wittink and Cattin 1989; Wittink, Vriens, and Burhenne 1994], provided extensive surveys attesting to the rapidity with which conjoint analysis was being adopted by researchers. Myers, Massy, and Greyser [1980] used conjoint analysis as a canonical case of how new ideas diffuse throughout the research community. Green and Srinivasan [1978] published a review of conjoint’s progress, which they followed with a further review 12 years later [Green and Srinivasan 1990]. The impetus behind conjoint’s diffusion reflected the joint influence of academic and practitioners’ contributions to the methodology, the availability of easy to use software, and the early credibility of results from business applications. Conjoint analysis, unlike MDS and clustering, dealt with central problems – measuring buyer trade-offs for developing new or reformulated products and estimating price-demand functions. In contrast, cluster analysis and MDS tools are often used as ancillary techniques for

11

data analysis and presentation. Both, however, have been gainfully employed in conjoint studies involving buyer segmentation [Green and Krieger 1991] and in presenting perceptual maps of conjoint results [Green, Krieger, and Carroll 1987]. Distant relatives to conjoint analysis include the decompositional approaches of Hoffman, Slovic, and Rorer [1968] and the functional measurement approach of Anderson [1970]. However, the seminal paper by Luce and Tukey [1964] and the experimental-design papers by Addelman [1962] and Plackett and Burman [1946], among others, provided key theoretical underpinnings and motivation for developing conjoint analysis. In particular, orthogonal main-effects designs and their more sophisticated extensions still constitute a major activity in the continuing development of efficient experimental designs. Considerable new research [Anderson and Wiley 1992; Kuhfeld, Tobias, and Garratt 1994; Lazari and Anderson 1994; Huber and Zwerina 1996] on this topic occurred during the ‘90s. The most publicized early paper on conjoint analysis appeared in Harvard Business Review [Green and Wind 1975]. Ironically, Green and Wind [1973] had earlier published a research monograph on conjoint analysis and multi-attribute decision making that met with much less reader interest.

Conjoint Development in the ‘80s Technical developments in conjoint analysis have proceeded swiftly over the past two decades, accompanied by thousands of applications (Table 2, next page). During the ‘80s, two developments stand out with regard to model development and application:

• Choice-based conjoint models, and • Hybrid conjoint models, including Johnson’s adaptive conjoint analysis model.

In traditional conjoint analysis, respondents typically rate various product or supplier profiles presented one at a time, on a likelihood-of purchase scale. In choice-based conjoint respondents typically see profile descriptions of two or more explicit competitors, which vary on one or more attributes. In this case, the task is either to pick one’s most preferred profile from the set or alternatively to allocate 100 points across the set of profiles, reflecting one’s relative strength of preferences. In choice-based conjoint analysis, analysts typically employ multinomial logit models, although occasionally they use probit-based models. They usually employ traditional conjoint analysis when a new product is entering a new or stable market in which competitors are either nonexistent or treated as “passive” in terms of responding competitively to the new entry. In choice-based conjoint analysis, however, the analyst assumes active competitors who can modify their profiles (including price and nonprice attributes) as well. Sawtooth Software’s Adaptive Conjoint Analysis is a traditional [hybrid] conjoint model. Its choice-based conjoint model deals with the case of explicit, active competitors. In choice-based conjoint analysis, analysts originally estimated parameters at the total sample level. Newer methods (hierarchical Bayes) now permit measurement of individual differences as well.

12

Table 2: A Partial List of Contributions to Conjoint Analysis over the 1974-2000 Period

Choice-based conjoint [McFadden 1974; Gensch and Recker 1979; Batsell and Lodish 1981; Mahajan, Green, and Goldberg 1982; Louviere and Woodworth 1983]

Three-way multivariate conjoint analysis [DeSarbo, Carroll, Lehmann, and O’Shaughnessy 1982]

Number of attribute levels effects on derived conjoint importance [Wittink, Krishnamurthi, and Nutter 1982]

Constrained parameter estimation in conjoint analysis [Srinivasan, Jain, and Malhotra 1983]

Hybrid models for conjoint analysis [Green, Goldberg , and Montemayor 1981; Green 1984]

Introduction in 1985 of Bretton-Clark’s full profile conjoint techniques [Herman 1988]

Introduction in 1985 of Sawtooth Software’s Adaptive Conjoint Analysis [Johnson 1987]

Factor analytic approaches to individualized conjoint analysis [Hagerty 1985]

Conjoint analysis and MDS in tandem [Green, Krieger, and Carroll 1987]

Reliability and validity testing [Bateson, Reibstein, and Boulding 1987]

Simultaneous conjoint parameter estimation and segmentation [Kamakura 1988]

Bretton-Clark’s second generation, full profile suite of programs [Herman 1988]

Conjunctive-compensatory self explicated models [Srinivasan 1988]

Componential segmentation with optimization features [Green, Krieger, and Zelnio 1989]

Compensatory model problems in negatively correlated environments [Johnson, Meyer, and Ghose 1989]

New experimental designs for conjoint [Steckel, DeSarbo, and Mahajan 1991]

A reservation price model for optimal pricing [Kohli and Mahajan 1991]

A review of experimental choice analysis [Batsell and Louviere 1991]

Latent class conjoint analysis [DeSarbo, Wedel, Vriens, and Ramaswamy 1992]

Constrained part-worth estimation [van der Lans and Heiser 1992]

Modeling hierarchical conjoint processes [Oppewal, Louviere and Timmermans 1994]

Concomitant variable latent class modeling [Kamakura, Wedel and Agrawal 1994]

Hierarchical Bayes models for conjoint analysis [Allenby, Arora, and Ginter 1995; Allenby and Ginter 1995; Lenk, DeSarbo, Green, and Young 1996]

A comparison of metric conjoint models [Vriens, Wedel and Wilms 1996]

Utility balanced experimental designs [Huber and Zwerina 1996]

Competitive interaction simulators [Choi, DeSarbo, and Harker 1990; Green and Krieger 1997]

Mixture models for segmentation [Wedel and Kamakura 1998]

Commercial windows-based, choice-based conjoint [Sawtooth Software 1999]

Krieger and Green’s hybrid choice-based conjoint model [see Vavra, Green and Krieger 1999]

Latent class conjoint analysis [Ramaswamy and Cohen 2000]

Hierarchical Bayes applied to Internet recommendation systems [Ansari, Essegaier, and Kohli 2000]

Response latencies and conjoint analysis [Haaijer, Kamakura, and Wedel 2000]

13

The seminal precursory paper to choice-based conjoint [using the multinomial logit model] was written by an econometrician [McFadden 1974]. McFadden’s work was soon recognized and adopted by a number of marketing researchers, including Punj and Staelin [1978]. Gensch and Recker [1979] also used this model in developing an alternative to regression for analyzing cross-sectional choice data. Batsell and Lodish [1981] illustrated the multinomial logit’s use in modeling individual choices over replicated choice sets. Their model yields a share-of-choice prediction for alternatives in competitive choice sets. They indicate that the model can be profitably applied to market segments, as well. Mahajan, Green and Goldberg [1982] applied Theil’s logit model to a choice-based conjoint problem. Louviere and Woodworth [1983] extended the preceding research. They discuss experimental designs that lend themselves to choice-based conjoint problems. They also provide an extensive and rich set of empirical examples based on various sets of fractional factorial designs. They focus on aggregate (pooled-over individuals) consumer choice studies. This paper spawned subsequent research for dealing with the more complex experimental designs needed for choice-based conjoint analysis. Choice-based conjoint studies can be a mixed blessing. The respondent’s tasks are extensive, since respondents may have to evaluate 10 (or more) scenarios. Each scenario could contain eight or more brands, each with several attributes and with several levels within attributes. Nonetheless, choice-based conjoint analysis has markedly increased in popularity because it can deal with the complexity of choosing among two or more competitive profiles, each of which can vary idiosyncratically across attributes and levels In the early ‘80s, hybrid models [Green, Goldberg, and Montemayor 1981; Green 1984] appeared in direct response to the increasing popularity of conjoint analysis. Along with conjoint’s increased application came the desire to expand the number of attributes and the levels that could be accommodated. Hybrid models employ self-explicated data collected on both the desirability and the importance of attributes and levels. Respondents then consider small subsets of the full profiles to evaluate. The early hybrid models initially used stagewise regression to fit simplified compositional models to the self-explicated data, later augmented by decompositional models [fitted at the segment level]. The value of these models lies in the greater accuracy they achieve, compared to non-hybrid models, in within-attribute estimation. The full-profile responses mainly serve to refine self-explicated attribute importances. More recently, Green and Krieger [1996] extended hybrid models to allow parameter estimation at the individual level. They describe four separate models of increasing generality. Again, the objective is to use the self-explicated data primarily for within-attribute part-worth estimation while using the full-profile analysis to produce improved estimates of attribute importances. The most used commercial conjoint model is Johnson’s adaptive-conjoint-analysis program. ACA is a hybrid model that incorporates self-explicated desirabilities and importances,

14

followed by the presentation of pairs of partial profiles (typically consisting of levels on two or perhaps three attributes) drawn from the full set of attributes. A respondent is asked to choose between the members of each pair and to include his or her preference intensity as well via graded paired comparison; ACA developers [Sawtooth Software 1999] have continued to introduce useful refinements to the original version [Johnson 1987].

Conjoint Developments in the ‘90s Probably the most far-reaching developments in the ‘90s use hierarchical Bayesian modeling of individual differences in choice-based models. Before this, choice based models were either estimated from data pooled across all individuals or by latent class methods (partial dissaggregation) as applied by DeSarbo, Wedel, Vriens, and Ramaswamy [1992] and Ramaswamy and Cohen [2000]. The work of Allenby, Arora, and Ginter [1995], Allenby and Ginter [1995], and Lenk, DeSarbo, Green, and Young [1996] has enabled choice-based conjoint users to obtain individual-level, part-worth estimates based on hierarchical Bayesian methods. (Sawtooth Software has recently added this type of module to its choice-based conjoint software.) To the extent that an individual’s parameters are both self-consistent and different from the aggregated data, the individual’s data will receive more weight in the estimation of his or her part-worths. Individuals whose part-worths are estimated poorly (that is, with large error in his or her own data) will receive more weight from the aggregate data. Recently, Vavra, Green and Krieger [1999] proposed another approach to the choice-based-conjoint problem. They developed a hybrid choice-based model and simulator that combines self-explicated data with full-profile responses. This model requires no weighted averaging of an individual’s data with that from the entire group. On the downside, however, it requires self-explicated [in addition to full profile] data.

Illustrative Applications Of the extensive list (Table 3, next page) of conjoint applications, two are of particular interest: the customer-driven design of Marriott’s Courtyard Hotels [Wind, Green, Shifflet, and Scarbrough 1989], and the design and evaluation of the New Jersey and New York EZ Pass electronic toll collection system [Vavra, Green, and Krieger 1999]. To the best of our knowledge, Courtyard by Marriott is the largest project in terms of attributes and levels ever undertaken using conjoint analysis. In the early ‘80s, Marriott management wished to design an “optimal” hotel chain catering primarily to business travelers who had no need for many of the features provided by up-scale hotels, such as Marriott and Hyatt.

15

Table 3: Illustrative Conjoint Applications AT&T’s first cellular telephone – Chicago based study of 1,000 drivers’ reactions to cell phone features of the new “honey-comb” relay system

The Ford Fairlane – conjoint analysis used in redesign of Ford to reflect general automotive downsizing objectives

IBM RISC 6000 workstation – conjoint used to measure potential buyer reactions to variations in performance and reliability features of a new workstation

Squibb’s captopril antihypertensive – six country study of physician’s evaluations of Capoten’s efficacy and safety features

Tagamet [SKF] and Zantac [Glaxo] ulcer drugs – competitive pricing and analysis of demand elasticities

Health maintenance plans – study conducted by the American Association of Retired Persons; results were submitted to Congress U.S. Navy benefit packages for re-enlistment – conjoint was used to develop menu of new re-enlistment plans, based on individual differences in types of duties, health needs, and sign-over bonuses Anti-dumping litigation – AT&T vs. Pacific rim manufacturers legal dispute regarding small business telephone equipment; case adjudicated in AT&T’s favor

Continental vs. American Airlines litigation – conjoint study of travel agents’ trade-offs among airline flight selections

FEDEX new services study –trade-off study of customer reaction to new methods for tracking delayed/lost letters and packages

Intermittent windshield wipers litigation – study to determine consumer evaluations of the derived “willingness to pay” for the intermittent wiper feature

MasterCard and Diner’s Club – new travel and entertainment features evaluated for each case

Monsanto’s herbicide packaging study – consumer reactions to advanced packaging devices for liquids, solids, and aerosols

Polaroid’s instant camera design – study of consumer reactions to new features and camera design aesthetics

Marriott time share units – development of optimal interior and exterior decors, services and price

Ritz Carlton – conjoint used to develop hotel décor and services

Japanese cable TV – conjoint survey of Japanese consumers’ trade-offs among services and prices of satellite TV

Brand equity – conjoint study of price, brand, and alternative company positions for the Shell Corporation

UPS services study – conjoint used to examine customers’ evaluations of four major suppliers of overnight letter/package delivery

In addition to price, the study design included seven hotel facets – external decor, room decor, food service, lounge facilities, general services, leisure activities (for example, a fitness club), and security features. Analysts developed some 50 attributes with a total of 160 attribute levels. The models included hybrid conjoint and an early type of choice-based conjoint. The analysts also used computer simulators. The study made extensive use of visual props (pictures

16

and three-dimensional models) as well as experimental rooms in which the furnishings and decor were systematically varied according to experimental designs. By all counts, the Courtyard study was a success. Marriott implemented almost all of the design recommendations, and later extended the approach to other new products (for example, Marriott Suites] and used the findings in the design of Courtyard advertisements and brochures. By 1990, Marriott had over 300 Courtyards, employing over 15,000 people. Today, there are 450 Courtyards worldwide, with annual sales in the billions of dollars. Marriott has since used conjoint analysis in such related endeavors as designing time-share vacation units and in room and amenities pricing. In the EZ-Pass toll collection project, New Jersey and New York developed a really new product aimed at speeding up and simplifying vehicle passage on their toll highways, bridges, and tunnels. Commuters use an electronic tag (transponder) attached to the vehicle’s inside windshield. As the vehicle approaches a toll lane, an antenna in the lane reads the customer’s vehicle and account information embedded in the tag. The information is electronically forwarded to an in-line computer. The computer, in turn, deducts the toll from the customer’s account. The project began in 1992. There were two main questions:

• How should EZ-Pass be configured? • What level of resources should be allocated to its implementation?

The two states conducted a large conjoint study (over 3,000 respondents) by a telephone-mail- telephone procedure in which qualified respondents received a packet of questionnaire materials and a video cassette. The video contained an 11-minute infomercial that described the problems associated with the current toll road system and the benefits provided by EZ-Pass. Analysts used seven conjoint attributes dealing with such issues as number of lanes available, tag acquisition, cost, toll prices, invoicing, and other uses of the tag. The study designers analyzed the individual respondents’ data at both the overall sample level and by region and facility. The overall (at equilibrium) forecast made in 1992 of “take rate” was 49 percent usage. The actual take rate (seven years later) was 44 percent; future usage is expected to be higher than 49 percent. (Even David Letterman liked the system!) Both projects, Courtyard by Marriott and EZ-Pass, illustrate the ability of conjoint analyses to lead to actionable findings that provide customer-driven design features and consumer usage or sales forecasts.

Future Prospects After 30 years of development and application, conjoint analysis seems to have survived the test of time. While new breakthroughs may well be less frequent, the method continues to grow in depth and breadth of usage. We expect to see the following further developments:

17

New simulator-optimizers that can maximize either financial return or market share [Vavra, Green, and Krieger 1999]; New classes of problems, including menu selection and bundling models for telecommunications and banking services [Ben-Akiva and Gershenfeld 1998]; More realistic imagery for describing attribute levels, for example, using virtual reality displays; Continued extensions of conjoint applications to such fields as tourism, entertainment, health maintenance, gambling, and legal disputes; New application venues, such as conjoint’s recent implementation on the Internet, including activebuyersguide.com, personalogic.com, and conjointonline.com, sites that typically use hybrid conjoint models to elicit buyer preferences for Web-based merchandise [Ansari, Essegaier, and Kohl 2000]; Additional studies of conjoint reliability and validity [Vriens, Wedel, and Wilms 1996; Haaijer, Kamakura, and Wedel 2000]; Extension of consumer-based applications to other stakeholder groups, such as employees, suppliers, stockholders, and municipalities; New “dynamic” conjoint simulators that consider competitive action-reaction sequences [Choi, DeSarbo, and Harker 1990; Green and Krieger 1997]; and Prototype simulators (for example, test cars) that permit analysts to measure respondents’ preferences for type of ride, acceleration, cornering, and so forth in realistic surroundings.

In short, despite its maturity, conjoint analysis is still far from stagnant. Because the

methods deal with the pervasive problem of buyer preferences and choices, conjoint’s future promises continued development and application.

The Role of Software Development in the Diffusion of Conjoint Analysis Throughout the development of conjoint analysis and the precursory techniques of cluster analysis and MDS, the availability of inexpensive and easy-to-use software has been crucial to their dissemination. Early on, Bell Laboratories and the Marketing Science Institute played important roles in making mainframe software available to both academic and industry users. With the growth of the personal computer, Johnson’s Sawtooth Software and Herman’s Bretton-Clark companies provided affordable PC software to business users and academic versions for teaching purposes. Sawtooth Software has maintained contact with business and academia through its newsletters, annual meetings, and continued development of software [Sawtooth

18

1999] to implement new research ideas, such as choice-based models. The American Marketing Association’s annual Advanced Research Techniques Forum provides a meeting place for the fruitful exchange of ideas between academics and practitioners.

A Science vs. Engineering Postscript From its inception, conjoint analysis has drawn on ideas from mathematical psychologists, psychometricians, statisticians, econometricians, and operations researchers. These ideas concern experimental design, parameter estimation, descriptive model building, normative model building, and the comparative evaluation of various models’ reliability and validity. The method’s practical consequences attest to its value and staying power [Gustaffsson Hermann, and Huber 2000]. Through the interplay of theoretical contributions and practical applications, conjoint methodology continues to grow, as academics and practitioners learn useful things from each other.

Acknowledgments We thank Eric Bradlow and Wes Hutchinson of the Wharton School, an anonymous reviewer, and the editors for helpful comments on the paper’s initial draft.

19

References Addelman, Sidney 1962, “Orthogonal main-effect plans for asymmetrical factorial experiments,”

Technometrics, Vol. 4 ( ), pp. 21-46. Allenby, Greg M., Arora, Neeraj, and Ginter, James L. 1995, “Incorporating prior knowledge

into the analysis of conjoint studies,” Journal of Marketing Research, Vol. 32 (May), pp. 152-162.

Allenby, Greg M. and Ginter, James L. 1995, “Using extremes to design products and segment

markets,” Journal of Marketing Research, Vol. 32 (May), pp. 152-162. Anderson, D. A. and Wiley, James B. 1992, “Efficient choice set designs for estimating cross-

effects models,” Marketing Letters, Vol. 3 (October), pp. 357-70. Anderson, Norman H. 1970, “Functional measurement and psychophysical judgment,”

Psychological Review, Vol. 77 ( ) pp. 153-70. Ansari, Asim, Essegaier, Skander, and Kohli, Rajeev 2000, “Internet recommendation systems,”

Journal of Marketing Research, Vol. 37 (August), pp.363-375. Bateson, John E.G., Reibstein, David J. and Boulding, William 1987, “Conjoint analysis

reliability and validity: A framework for future research,” Review of Marketing, ed., M.J. Houston, American Marketing Association, Chicago, IL, pp. 451-481.

Batsell, Richard R. and Lodish, Leonard M. 1981, “A model and measurement methodology for

predicting individual consumer choice,” Journal of Marketing Research, Vol. 18 (February), pp. 1-12.

Batsell, Richard R. and Louviere, Jordan J. 1991, “Experimental analysis of choice,” Marketing

Letters, Vol. 2 (August), pp. 199-214. Bell, D. F., Raiffa, Howard and Tversky, Amos, eds. 1988, Decision Making: Descriptive,

Normative, and Prescriptive Interactions, Cambridge University Press, New York, NY. Ben-Akiva, Moshe and Gershenfeld, Shari 1998, “Multi-featured products and services:

Analyzing pricing and bundling strategies,” Journal of Forecasting, Vol. 17, pp. 175-196.

Carmone, Frank J., Green, Paul E., and Jain, Arun K. 1978, “The robustness of conjoint analysis:

Some Monte Carlo results,” Journal of Marketing Research, Vol. 15, pp. 300-303. Carroll, J. Douglas 1969, “Categorical conjoint measurement,” Meeting of Mathematical

Psychology, Ann Arbor, MI.

20

Carroll, J. Douglas and Green, Paul E. 1997, “Psychometric methods in marketing research: Part II, multidimensional scaling,” Journal of Marketing Research, Vol. 34 (May), pp. 193-204.

Cattin, Philippe and Wittink, Dick R. 1976, “A Monte Carlo study of metric and nonmetric

estimation techniques,” Paper 341, Graduate School of Business, Stanford University. Choi, S. C., DeSarbo, Wayne S., and Harker, Patrick T. 1990, “Product positioning under price

competition,” Management Science, Vol. 30, pp. 175-199. DeSarbo, Wayne S., Carroll, J. Douglas, Lehmann, Donald R., and O’Shaugnhessy, John 1982.

“Three-way multivariate conjoint analysis,” Marketing Science, Vol. 1 (Fall), pp. 323-350.

DeSarbo, Wayne S., Wedel, Michel, Vriens, Marco, and Ramaswamy, Venkatram 1992, “Latent

class metric conjoint analysis,” Marketing Letters, Vol. 3 (July), pp. 273-288. Gensch, Dennis H. and Recker, Wilfred W. 1979, “The multinominal multiattribute logit choice

model,” Journal of Marketing Research, Vol. 16 (February), pp. 124-132. Green, Paul E. 1984, “Hybrid models for conjoint analysis: An expository review,” Journal of

Marketing Research, Vol. 21 (May), pp. 155-159. Green, Paul E., Frank, Ronald E., and Robinson, Patrick J. 1967, “Cluster analysis in test market

selection,” Management Science, Vol. 13, pp. B387-400. Green, Paul E., Goldberg, Stephen M., and Montemayor, Mila 1981, “A hybrid utility estimation

model for conjoint analysis,” Journal of Marketing, Vol. 45 (Winter), pp. 33-41. Green, Paul E. and Krieger, Abba M. 1991, “Segmenting markets with conjoint analysis,”

Journal of Marketing, Vol. 55 (October), pp. 20-31. Green, Paul E. and Krieger, Abba M. 1996, “Individualized hybrid models for conjoint analysis,”

Management Science, Vol. 42 (June), pp. 850-867. Green, Paul E. and Krieger, Abba M. 1997, “Using conjoint analysis to view competitive

interaction through the customer’s eyes,” in Wharton on Dynamic Competitive Strategy, eds., George S. Day, Reibstein, D. J., and Gunther, R.E., John Wiley & Sons, New York, N.Y., pp. 343-367.

Green, Paul E., Krieger, Abba M., and Carroll, J. Douglas 1987, “Conjoint analysis and

multidimensional scaling: A complementary approach,” Journal of Advertising Research, Vol. 27 (October/November), pp. 21-27.

Green, Paul E., Krieger, Abba M., and Zelnio, Robert N. 1989, “A componential segmentation

model with optimal design features,” Decision Sciences, Vol. 20 (Spring), pp. 221-238.

21

Green, Paul E. and Rao, Vithala R. 1971, “Conjoint measurement for quantifying judgmental

data,” Journal of Marketing Research, Vol. 8 (August), pp. 355-363. Green, Paul E. and Srinivasan ,V. 1978, “Conjoint analysis in consumer research: issues and

outlook,” Journal of Consumer Research, Vol. 5 (September), pp. 103-123. Green, Paul E. and Srinivasan ,V. 1990, “Conjoint analysis in marketing: New developments

with implications for research and practice,” Journal of Marketing, Vol. 54 (October), pp. 3-19.

Green, Paul E. and Wind, Yoram 1973, Multiattribute Decisions in Marketing: A Measurement

Approach, The Dryden Press, Hinsdale, IL. Green, Paul E. and Wind, Yoram 1975, “New way to measure consumers’ judgments,” Harvard

Business Review, Vol. 53 (July-August), pp. 107-117. Gustaffsson, Anders, Hermann, Andress, and Huber, Frank, eds. 2000, Conjoint Measurement:

Methods and Applications, Springer-Verlag, Berlin, Germany. Haaijer, Rinus, Kamakura, Wagner, and Wedel, Michel 2000, “Response latencies in the analysis

of conjoint choice experiments,” Journal of Marketing Research, Vol. 37 (August), pp. 376-382.

Hagerty, Michael R. (1985), “Improving the predictive power of conjoint analysis: The use of

factor analysis and cluster analysis,” Journal of Marketing Research, Vol. 22 (May), 168-184.

Herman, Steve 1988, “Software for full-profile conjoint analysis,” in Proceedings of the

Sawtooth Conference on Perceptual Mapping, Conjoint Analysis, and Computer Interviewing, ed., M. Metegrano, Sawtooth Software, Ketchum, ID, pp. 117-130.

Hoffman, Paul J., Slovic, Paul and Rorer, L. G. 1968 “An analysis of variance model for the

assessment of configural cue utilization in clinical judgment,” Psychological Bulletin, Vol. 69, pp. 338-349.

Huber, Joel and Zwerina, Klaus 1996, “The importance of utility balance in efficient choice

designs,” Journal of Marketing Research, Vol. 33 (August), pp. 307-317. Johnson, Eric, Meyer, Robert J., and Ghose, Sanjay 1989, “When choice models fail:

Compensatory models in negatively correlated environments,” Journal of Marketing Research, Vol. 26 (August), pp. 255-270.

Johnson, Richard M. 1974, “Trade-off analysis of consumer values,” Journal of Marketing

Research, Vol. 11 (May), pp. 121-127.

22

Johnson, Richard M. 1987, “Adaptive conjoint analysis,” in Sawtooth Software Conference on Perceptual Mapping, Conjoint Analysis, and Computer Interviewing. Sawtooth Software, Ketchum, ID, pp. 253-265.

Kamakura, Wagner 1988, “A least squares procedure for benefit segmentation with conjoint

experiments,” Journal of Marketing Research, Vol. 25 (May), pp. 157-167. Kamakura, Wagner, Wedel, Michel, and Agrawal, Jagdish 1994, “Concomitant variable latent

class models for conjoint analysis,” International Journal of Research in Marketing, Vol. 11, pp. 451-464.

Kaul, Anil and Rao, Vithala R. 1994, “Research for product position and design decisions: An

integrative review,” International Journal of Research on Marketing, Vol. 12, pp. 293-320.

Keeney, Ralph L. and Raiffa, Howard 1976, Decisions with Multiple Objectives: Preferences

and Value Tradoffs, Wiley, New York, NY. Kohli, Rajeev and Mahajan, Vijay 1991, “A reservation price model for optimal pricing of

multiattribute products in conjoint analysis,” Journal of Marketing Research, Vol. 28 (August), pp. 347-354.

Kruskal, Joseph B. 1965, “Analysis of factorial experiments by estimating monotone

transformations of the data,” Journal of the Royal Statistical Society, Series B, Vol. 27, pp. 251-263.

Kuhfeld, Warren F., Tobias, Randall D., and Garratt, Mark 1994, “Efficient experimental designs

with marketing research applications,” Journal of Marketing Research, Vol. 31 (November), pp. 545-557.

Lazari, Andreas G. and Anderson, Donald A. 1994, “Designs of discrete choice set experiments

for estimating both attribute and availability cross effects,” Journal of Marketing Research, Vol. 31 (August), pp. 375-383.

Lenk, Peter J., DeSarbo, Wayne S., Green, Paul E., and Young, Martin R. 1996, “Hierarchical

Bayes conjoint analysis: Recovery of part-worth heterogeneity from reduced experimental designs,” Marketing Science, Vol. 15, No. 2, pp. 173-191.

Louviere, Jordan and Woodworth, George 1983, “Design and analysis of simulated consumer

choice or allocation experiments, Journal of Marketing Research, Vol. 20 (November), pp. 350-367.

Luce, R. Duncan and Tukey, John W. 1964, “Simultaneous conjoint measurement: A new type

of fundamental measurement,” Journal of Mathematical Psychology, Vol. 1, pp. 1-27.

23

Mahajan, Vijay, Green, Paul E., and Goldberg, Stephen M. 1982, “A conjoint model for measuring self-and cross-price demand relationships,” Journal of Marketing Research, Vol. 19 (August), pp. 334-342.

McFadden, Daniel 1974, “Conditional logit analysis of qualitative choice behavior,” in Frontiers

on Econometrics, ed., P. Zarembka, Academic Press, New York, NY, pp. 105-421. Myers, John G., Massy, William F., and Greyser, Stephen A. 1980, Marketing Research and

Knowledge Development. Prentice-Hall, Englewood Cliffs, NJ. Oppewal, Harmen, Louviere, Jordan, and Timmermans, Harry 1994, “Modeling hierarchical

conjoint processes with integrated choice experiments,” Journal of Marketing Research, Vol. 31 (February), pp. 92-105.

Plackett, R. L. and Burman, J. P. 1946, “The design of optimum multifactorial experiments,”

Biometrika, Vol. 33, pp. 305-325. Punj, Girish N. and Staelin, Richard 1978, “The choice process for graduate business schools,”

Journal of Marketing Research, Vol. 15 (November), pp. 588-598. Ramaswamy, Venkatram and Cohen, Steven H. 2000, “Latent class models for conjoint

analysis,” in Conjoint Measurement: Methods and Applications, eds., A. Gustafsson, A. Hermann, and F. Huber, Springer-Verlag, Berlin, Germany.

Saaty, Thomas L. 1980, The Analytical Hierarchy Process, McGraw-Hill, New York, NY. Sawtooth Software 1999, CBC for Windows, Sequim, WA. Shocker, Allan D., and Srinivasan, V. 1977, “LINMAP (Version II): A FORTRAN IV computer

program for analyzing ordinal preference (dominance) judgments via linear programming techniques and for conjoint measurement,” Journal of Marketing Research, Vol. 14, pp. 101-103.

Srinivasan, V. 1988, “A conjunctive-compensatory approach to the self-explication of

multiattributed preferences,” Decision Sciences, Vol. 19 (Spring), pp. 295-305. Srinivasan, V., Jain, Arun K., and Malhotra, Naresh K. 1983, “Improving the predictive power of

conjoint analysis by constrained parameter estimation,” Journal of Marketing Research, Vol. 20 (November), pp. 433-438.

Starr, Martin K. and Zeleny, M. 1977, Multiple Criteria Decisions Making, North-Holland,

Amsterdam, Holland. Steckel, Joel H., DeSarbo, Wayne S., and Mahajan, Vijay 1991, “On the creation of feasible

conjoint analysis experimental designs,” Decision Sciences, Vol. 22, pp. 435-442.

24

van der Lans, Ivo A. and Heiser, Willem H. 1992. “Constrained part-worth estimation in conjoint analysis using the self-explicated utility model,” International Journal of Research in Marketing, Vol. 9, pp. 325-344.

Vavra, Terry G., Green, Paul E., and Krieger, Abba M. 1999, “Evaluating EZ-Pass,” Marketing

Research, Vol. 11 (Summer), pp. 5-16. Vriens, Marco, Wedel, Michel, and Wilms, Tom 1996, “Metric conjoint segmentation methods:

A Monte Carlo comparison, Journal of Marketing Research, Vol. 33 (February), pp. 73-85.

Wedel, Michel and Kamakura, Wagner A. 1998, Market Segmentation: Conceptual and

Methodological Foundations, Kluwer, Boston, MA. Westwood, Dick, Lunn, Tony, and Beazley, David 1974, “The trade-off model and its

extensions,” Journal of the Market Research Society, Vol. 16, pp. 227-241. Wind, Jerry, Green, Paul E., Shifflet, Douglas, and Scarbrough, Marsha 1989 “Courtyard by

Marriott: Designing a hotel facility with consumer-based marketing models,” Interfaces Vol. 19 (January-February), pp. 25-47.

Wittink, Dick and Cattin, Philippe 1989, “Commercial use of conjoint analysis: An update,”

Journal of Marketing, Vol. 53 (July), pp. 91-96. Wittink, Dick R., Krishnamurthi, Lakshman, and Nutter, Julia B. 1982, “Comparing derived

importance weights across attributes,” Journal of Consumer Research, Vol. 8 (March), pp. 471-474.

Wittink, Dick, Vriens, Marco, and Burhenne, Wim 1994, “Commercial use of conjoint in europe:

Results and critical reflections,” International Journal of Research in Marketing, Vol. 11, pp. 41-52.

Young, Forrest W. 1969, “Polynomial conjoint analysis of similarities: Definitions for a special

algorithm,” Research paper No. 76, Psychometric Laboratory, University of North Carolina.